Python chainer.functions 模块，sqrt() 实例源码

def __call__(self, loc, val, y, train=True):
        bs = val.data.shape[0]
        pred, kld0, kld1, kld2 = self.forward(loc, val, y, train=train)

        # Compute MSE loss
        mse = F.mean_squared_error(pred, y)
        rmse = F.sqrt(mse)  # Only used for reporting

        # Now compute the total KLD loss
        kldt = kld0 * self.lambda0 + kld1 * self.lambda1 + kld2 * self.lambda2

        # Total loss is MSE plus regularization losses
        loss = mse + kldt * (1.0 / self.total_nobs)

        # Log the errors
        logs = {'loss': loss, 'rmse': rmse, 'kld0': kld0, 'kld1': kld1,
                'kld2': kld2, 'kldt': kldt, 'bias': F.sum(self.bias_mu.b)}
        reporter.report(logs, self)
        return loss

def __call__(self, loc, val, y, train=True):
        bs = val.data.shape[0]
        ret = self.forward(loc, val, y, train=train)
        pred, kld0, kld1, kldg, kldi, hypg, hypi = ret

        # Compute MSE loss
        mse = F.mean_squared_error(pred, y)
        rmse = F.sqrt(mse)  # Only used for reporting

        # Now compute the total KLD loss
        kldt = kld0 * self.lambda0 + kld1 * self.lambda1
        kldt += kldg + kldi + hypg + hypi

        # Total loss is MSE plus regularization losses
        loss = mse + kldt * (1.0 / self.total_nobs)

        # Log the errors
        logs = {'loss': loss, 'rmse': rmse, 'kld0': kld0, 'kld1': kld1,
                'kldg': kldg, 'kldi': kldi, 'hypg': hypg, 'hypi': hypi,
                'hypglv': F.sum(self.hyper_feat_lv_vec.b),
                'hypilv': F.sum(self.hyper_feat_delta_lv.b),
                'kldt': kldt, 'bias': F.sum(self.bias_mu.b)}
        reporter.report(logs, self)
        return loss

def instance_norm(self, x, gamma=None, beta=None):
        mean = F.mean(x, axis=-1)
        mean = F.mean(mean, axis=-1)
        mean = F.broadcast_to(mean[Ellipsis, None, None], x.shape)
        var = F.squared_difference(x, mean)
        std = F.sqrt(var + 1e-5)
        x_hat = (x - mean) / std
        if gamma is not None:
            gamma = F.broadcast_to(gamma[None, Ellipsis, None, None], x.shape)
            beta = F.broadcast_to(beta[None, Ellipsis, None, None], x.shape)
            return gamma * x_hat + beta
        else:
            return x_hat

def lifted_struct_loss(f_a, f_p, alpha=1.0):
    """Lifted struct loss function.

    Args:
        f_a (~chainer.Variable): Feature vectors as anchor examples.
            All examples must be different classes each other.
        f_p (~chainer.Variable): Positive examples corresponding to f_a.
            Each example must be the same class for each example in f_a.
        alpha (~float): The margin parameter.

    Returns:
        ~chainer.Variable: Loss value.

    See: `Deep Metric Learning via Lifted Structured Feature Embedding \
        <http://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/\
        Song_Deep_Metric_Learning_CVPR_2016_paper.pdf>`_

    """
    assert f_a.shape == f_p.shape, 'f_a and f_p must have same shape.'
    n = 2 * f_a.shape[0]  # use shape[0] due to len(Variable) returns its size
    f = F.vstack((f_a, f_p))
    D_sq = squared_distance_matrix(f)

    pairs_p = np.arange(n).reshape(2, -1)  # indexes of positive pairs
    row = []
    col = []
    for i, j in pairs_p.T:
        row.append([i] * (n - 2) + [j] * (n - 2))
        col.append(np.tile(np.delete(np.arange(n), (i, j)), 2))
    row = np.ravel(row)
    col = np.ravel(col)
    pairs_n = np.vstack((row, col))

    distances_p = F.sqrt(D_sq[pairs_p[0], pairs_p[1]])
    distances_n = F.sqrt(D_sq[pairs_n[0], pairs_n[1]])
    distances_n = distances_n.reshape((n // 2, -1))
    loss_ij = F.logsumexp(alpha - distances_n, axis=1) + distances_p
    return F.sum(F.relu(loss_ij) ** 2) / n

def __init__(self, n_features=None, n_dim=8, lossfun=F.mean_squared_error,
                 lambda0=1, lambda1=1, lambda2=1, init_bias_mu=0.0,
                 init_bias_lv=0.0, intx_term=True, total_nobs=1):
        self.n_dim = n_dim
        self.n_features = n_features
        self.lossfun = lossfun
        self.lambda0 = lambda0
        self.lambda1 = lambda1
        self.lambda2 = lambda2
        self.intx_term = intx_term
        self.total_nobs = total_nobs

        # In contrast to the FM model, the slopes and latent vectors
        # will have means (mu) and log variances (lv) for each component.
        super(VFM, self).__init__(bias_mu=L.Bias(shape=(1,)),
                                  bias_lv=L.Bias(shape=(1,)),
                                  slop_mu=L.Bias(shape=(1, 1)),
                                  slop_lv=L.Bias(shape=(1, 1)),
                                  slop_delta_mu=L.EmbedID(n_features, 1,
                                                          ignore_label=-1),
                                  slop_delta_lv=L.EmbedID(n_features, 1,
                                                          ignore_label=-1),
                                  feat_mu_vec=L.Bias(shape=(1, 1, n_dim)),
                                  feat_lv_vec=L.Bias(shape=(1, 1, n_dim)),
                                  feat_delta_mu=L.EmbedID(n_features, n_dim,
                                                          ignore_label=-1),
                                  feat_delta_lv=L.EmbedID(n_features, n_dim,
                                                          ignore_label=-1))

        # Xavier initialize weights
        c = np.sqrt(n_features * n_dim) * 1e3
        d = np.sqrt(n_features) * 1e3
        self.feat_delta_mu.W.data[...] = np.random.randn(n_features, n_dim) / c
        self.feat_delta_lv.W.data[...] = np.random.randn(n_features, n_dim) / c
        self.slop_delta_mu.W.data[...] = np.random.randn(n_features, 1) / d
        self.slop_delta_lv.W.data[...] = np.random.randn(n_features, 1) / d
        self.bias_mu.b.data[...] *= 0.0
        self.bias_mu.b.data[...] += init_bias_mu
        self.bias_lv.b.data[...] *= 0.0
        self.bias_lv.b.data[...] += init_bias_lv

def __init__(self, n_features=None, n_dim=8, lossfun=F.mean_squared_error,
                 lambda0=5e-3, lambda1=5e-3, lambda2=5e-3, init_bias=0.0,
                 intx_term=True, total_nobs=1):
        self.n_dim = n_dim
        self.n_features = n_features
        self.lossfun = lossfun
        self.lambda0 = lambda0
        self.lambda1 = lambda1
        self.lambda2 = lambda2
        self.intx_term = intx_term
        self.total_nobs = total_nobs

        # These are all the learned weights corresponding
        # to the overall bias, slope per feature, and latent
        # interaction vector per feature
        super(FM, self).__init__(bias=L.Bias(shape=(1,)),
                                 slope=L.EmbedID(n_features, 1),
                                 latent=L.EmbedID(n_features, n_dim))

        # Xavier initialize weights
        c = np.sqrt(n_features * n_dim)
        self.latent.W.data[...] = np.random.randn(n_features, n_dim) / c
        d = np.sqrt(n_features)
        self.slope.W.data[...] = np.random.randn(n_features, 1) / d
        self.bias.b.data[...] *= 0.0
        self.bias.b.data[...] += init_bias

def average_loss(self, h, a, t):
        ## print F.reshape(t, (-1, 1)).data        
        ## print (h-F.reshape(t, (-1, 1))).data
        self.loss = F.sum(abs(h - F.reshape(t, (-1,1)))) 
        ## self.loss = F.sqrt(F.sum(F.square(h - F.reshape(t, (-1,1))))) 
        self.loss /= self.n_patches
        if self.n_images > 1:
            h = F.split_axis(h, self.n_images, 0)
            a = F.split_axis(a, self.n_images, 0)
        else:
            h, a = [h], [a]

        self.y = h
        self.a = a

def update_core(self):
        xp = self.gen.xp
        self._iter += 1

        opt_d = self.get_optimizer('dis')
        for i in range(self._dis_iter):
            d_fake = self.get_fake_image_batch()
            d_real = self.get_real_image_batch()

            y_fake = self.dis(Variable(d_fake), test=False)
            y_real = self.dis(Variable(d_real), test=False)

            w1 = F.average(y_fake-y_real)

            loss_dis = w1

            if self._mode == 'gp':
                eta = np.random.rand()
                c = (d_real * eta + (1.0 - eta) * d_fake).astype('f')
                y = self.dis(Variable(c), test=False, retain_forward=True)

                g = xp.ones_like(y.data)
                grad_c = self.dis.differentiable_backward(Variable(g))
                grad_c_l2 = F.sqrt(F.sum(grad_c**2, axis=(1, 2, 3)))

                loss_gp = loss_l2(grad_c_l2, 1.0)

                loss_dis += self._lambda_gp * loss_gp

            opt_d.zero_grads()
            loss_dis.backward()
            opt_d.update()

            if self._mode == 'clip':
                self.dis.clip()

        chainer.report({'loss': loss_dis,'loss_w1': w1}, self.dis)

        z_in = self.get_latent_code_batch()
        x_out = self.gen(Variable(z_in), test=False)

        opt_g = self.get_optimizer('gen')
        y_fake = self.dis(x_out, test=False)
        loss_gen = - F.average(y_fake)

        chainer.report({'loss': loss_gen}, self.gen)

        opt_g.zero_grads()
        loss_gen.backward()
        opt_g.update()

def update_core(self):
        xp = self.gen.xp
        self._iter += 1

        opt_g = self.get_optimizer('gen')
        opt_d = self.get_optimizer('dis')

        data_z0 = self.get_latent_code_batch()
        x_fake0 = self.gen(Variable(data_z0))
        data_z1 = self.get_latent_code_batch()
        x_fake1 = self.gen(Variable(data_z1))
        data_x = self.get_real_image_batch()
        x_real = Variable(data_x)

        eta = np.random.rand()
        x_inter = Variable((data_x * eta + (1.0 - eta) * x_fake0.data).astype('f'))

        dis_x_fake0 = self.dis(x_fake0)
        dis_x_fake1 = self.dis(x_fake1)
        dis_x_real = self.dis(x_real)

        loss_gen = loss_l2_norm(dis_x_fake0, dis_x_real) + \
                    loss_l2_norm(dis_x_fake1, dis_x_real) - \
                    loss_l2_norm(dis_x_fake0, dis_x_fake1)
        #print(loss_gen.data)

        chainer.report({'loss': loss_gen}, self.gen)
        opt_g.zero_grads()
        loss_gen.backward()
        opt_g.update()

        x_fake0.unchain_backward()
        x_fake1.unchain_backward()

        loss_surrogate = loss_l2_norm(dis_x_fake0, dis_x_fake1) - \
                    loss_l2_norm(dis_x_fake0, 0.0) + \
                    loss_l2_norm(dis_x_real, 0.0) - \
                    loss_l2_norm(dis_x_real, dis_x_fake1)

        dis_x_inter = self.dis(x_inter, retain_forward=True)
        g = xp.ones_like(dis_x_inter.data)
        t0 = dis_x_inter.data - dis_x_fake1.data
        t0_norm = xp.sum(t0**2, axis=(1)) ** 0.5
        t1_norm = xp.sum(dis_x_inter.data**2, axis=(1)) ** 0.5
        t_g = ((t0.transpose() / t0_norm) - (dis_x_inter.data.transpose()) / t1_norm).transpose()
        g = g * t_g

        grad = self.dis.differentiable_backward(Variable(g))
        grad_l2 = F.sqrt(F.sum(grad**2, axis=(1, 2, 3)))
        loss_gp = self._lambda_gp * loss_l2(grad_l2, 1.0)

        loss_dis = loss_surrogate + loss_gp

        opt_d.zero_grads()
        loss_dis.backward()
        opt_d.update()

        chainer.report({'loss': loss_dis, 'loss_gp': loss_gp}, self.dis)

def update_core(self):
        xp = self.gen.xp
        self._iter += 1

        opt_g = self.get_optimizer('gen')
        opt_d = self.get_optimizer('dis')

        data_z = self.get_latent_code_batch()
        data_x = self.get_real_image_batch()

        x_fake = self.gen(Variable(data_z))



        dis_fake = self.dis(x_fake)

        loss_gen = loss_func_dcgan_dis_real(dis_fake)
        chainer.report({'loss': loss_gen}, self.gen)
        opt_g.zero_grads()
        loss_gen.backward()
        opt_g.update()

        x_fake.unchain_backward()

        std_data_x = xp.std(data_x, axis=0, keepdims=True)
        rnd_x = xp.random.uniform(0, 1, data_x.shape).astype("f")
        x_perturbed = Variable(data_x + 0.5*rnd_x*std_data_x)

        x_real = Variable(data_x)
        dis_real = self.dis(x_real)
        dis_perturbed = self.dis(x_perturbed, retain_forward=True)
        g = Variable(xp.ones_like(dis_perturbed.data))
        grad = self.dis.differentiable_backward(g)

        grad_l2 = F.sqrt(F.sum(grad**2, axis=(1, 2, 3)))
        loss_gp = self._lambda_gp * loss_l2(grad_l2, 1.0)

        loss_dis = loss_func_dcgan_dis_real(dis_real) + \
                    loss_func_dcgan_dis_fake(dis_fake) + \
                    loss_gp

        opt_d.zero_grads()
        loss_dis.backward()
        opt_d.update()

        chainer.report({'loss': loss_dis, 'loss_gp': loss_gp}, self.dis)

def update_core(self):
        xp = self.gen.xp
        self._iter += 1

        opt_g = self.get_optimizer('gen')
        opt_d = self.get_optimizer('dis')

        data_z = self.get_latent_code_batch()
        data_tag = self.get_fake_tag_batch()

        data_x, data_real_tag = self.get_real_image_batch()

        x_fake = self.gen(F.concat([Variable(data_z),Variable(data_tag)]))

        dis_fake, dis_g_class = self.dis(x_fake)
        data_tag[data_tag < 0] = 0.0
        loss_g_class =loss_sigmoid_cross_entropy_with_logits(dis_g_class, data_tag)
        #print(loss_g_class.data)
        loss_gen = self._lambda_adv * loss_func_dcgan_dis_real(dis_fake) + loss_g_class
        chainer.report({'loss': loss_gen, 'loss_c': loss_g_class}, self.gen)

        opt_g.zero_grads()
        loss_gen.backward()
        opt_g.update()

        x_fake.unchain_backward()

        std_data_x = xp.std(data_x, axis=0, keepdims=True)
        rnd_x = xp.random.uniform(0, 1, data_x.shape).astype("f")
        x_perturbed = Variable(data_x + 0.5*rnd_x*std_data_x)

        x_real = Variable(data_x)
        dis_real, dis_d_class = self.dis(x_real)
        dis_perturbed, _ = self.dis(x_perturbed, retain_forward=True)
        g = Variable(xp.ones_like(dis_perturbed.data))
        grad = self.dis.differentiable_backward(g)

        grad_l2 = F.sqrt(F.sum(grad**2, axis=(1, 2, 3)))
        loss_gp = self._lambda_gp * loss_l2(grad_l2, 1.0)

        loss_d_class = loss_sigmoid_cross_entropy_with_logits(dis_d_class, data_real_tag)

        loss_dis = self._lambda_adv * ( loss_func_dcgan_dis_real(dis_real) + \
                    loss_func_dcgan_dis_fake(dis_fake) )+ \
                    loss_d_class + \
                    loss_gp

        opt_d.zero_grads()
        loss_dis.backward()
        opt_d.update()

        chainer.report({'loss': loss_dis, 'loss_gp': loss_gp, 'loss_c': loss_d_class}, self.dis)

def compute_distance_of_cluster_heads(self):
        # list all possible combinations of two cluster heads
        num_combination = self.nCr(self.ndim_y, 2)

        # a_labels
        # [0, 1, 0, 0]
        # [0, 0, 1, 0]
        # [0, 0, 1, 0]
        # [0, 0, 0, 1]
        # [0, 0, 0, 1]
        # [0, 0, 0, 1]
        a_labels = np.zeros((num_combination, self.ndim_y), dtype=np.float32)
        for i in range(1, self.ndim_y):
            for n in range(i):
                j = int(0.5 * i * (i - 1) + n)
                a_labels[j, i] = 1

        # b_labels
        # [1, 0, 0, 0]
        # [1, 0, 0, 0]
        # [0, 1, 0, 0]
        # [1, 0, 0, 0]
        # [0, 1, 0, 0]
        # [0, 0, 1, 0]
        b_labels = np.zeros((num_combination, self.ndim_y), dtype=np.float32)
        for i in range(1, self.ndim_y):
            for n in range(i):
                j = int(0.5 * i * (i - 1) + n)
                b_labels[j, n] = 1


        xp = self.xp
        if xp is not np:
            a_labels = cuda.to_gpu(a_labels)
            b_labels = cuda.to_gpu(b_labels)

        a_vector = self.cluster_head(a_labels)
        b_vector = self.cluster_head(b_labels)
        distance = functions.sqrt(functions.sum((a_vector - b_vector) ** 2, axis=1))

        # clip
        distance = functions.clip(distance, 0.0, float(self.cluster_head_distance_threshold))

        return distance

def compute_distance_of_cluster_heads(self):
        # list all possible combinations of two cluster heads
        num_combination = self.nCr(self.ndim_y, 2)

        # a_labels
        # [0, 1, 0, 0]
        # [0, 0, 1, 0]
        # [0, 0, 1, 0]
        # [0, 0, 0, 1]
        # [0, 0, 0, 1]
        # [0, 0, 0, 1]
        a_labels = np.zeros((num_combination, self.ndim_y), dtype=np.float32)
        for i in range(1, self.ndim_y):
            for n in range(i):
                j = int(0.5 * i * (i - 1) + n)
                a_labels[j, i] = 1

        # b_labels
        # [1, 0, 0, 0]
        # [1, 0, 0, 0]
        # [0, 1, 0, 0]
        # [1, 0, 0, 0]
        # [0, 1, 0, 0]
        # [0, 0, 1, 0]
        b_labels = np.zeros((num_combination, self.ndim_y), dtype=np.float32)
        for i in range(1, self.ndim_y):
            for n in range(i):
                j = int(0.5 * i * (i - 1) + n)
                b_labels[j, n] = 1


        xp = self.xp
        if xp is not np:
            a_labels = cuda.to_gpu(a_labels)
            b_labels = cuda.to_gpu(b_labels)

        a_vector = self.cluster_head(a_labels)
        b_vector = self.cluster_head(b_labels)
        distance = functions.sqrt(functions.sum((a_vector - b_vector) ** 2, axis=1))

        # clip
        distance = functions.clip(distance, 0.0, float(self.cluster_head_distance_threshold))

        return distance

def __init__(self, n_features=None, n_dim=8, lossfun=F.mean_squared_error,
                 lambda0=1, lambda1=1, lambda2=1, init_bias_mu=0.0,
                 init_bias_lv=0.0, intx_term=True, total_nobs=1):
        self.n_dim = n_dim
        self.n_features = n_features
        self.lossfun = lossfun
        self.lambda0 = lambda0
        self.lambda1 = lambda1
        self.lambda2 = lambda2
        self.intx_term = intx_term
        self.total_nobs = total_nobs

        # In contrast to the FM model, the slopes and latent vectors
        # will have means (mu) and log variances (lv) for each component.
        ones_3d = (1, 1, 1)
        super(AutoVFM, self).__init__(bias_mu=L.Bias(shape=(1,)),
                                      bias_lv=L.Bias(shape=(1,)),
                                      slop_mu=L.Bias(shape=(1, 1)),
                                      slop_lv=L.Bias(shape=(1, 1)),
                                      slop_delta_mu=L.EmbedID(n_features, 1,
                                                              ignore_label=-1),
                                      slop_delta_lv=L.EmbedID(n_features, 1,
                                                              ignore_label=-1),
                                      feat_mu_vec=L.Bias(shape=(1, 1, n_dim)),
                                      feat_lv_vec=L.Bias(shape=(1, 1, n_dim)),
                                      hyper_feat_lv_vec=L.Bias(shape=ones_3d),
                                      feat_delta_mu=L.EmbedID(n_features, n_dim,
                                                              ignore_label=-1),
                                      feat_delta_lv=L.EmbedID(n_features, n_dim,
                                                              ignore_label=-1),
                                      hyper_feat_delta_lv=L.Bias(shape=ones_3d))

        # Xavier initialize weights
        c = np.sqrt(n_features * n_dim) * 1e3
        d = np.sqrt(n_features) * 1e3
        self.feat_delta_mu.W.data[...] = np.random.randn(n_features, n_dim) / c
        self.feat_delta_lv.W.data[...] = np.random.randn(n_features, n_dim) / c
        self.slop_delta_mu.W.data[...] = np.random.randn(n_features, 1) / d
        self.slop_delta_lv.W.data[...] = np.random.randn(n_features, 1) / d
        self.bias_mu.b.data[...] *= 0.0
        self.bias_mu.b.data[...] += init_bias_mu
        self.bias_lv.b.data[...] *= 0.0
        self.bias_lv.b.data[...] += init_bias_lv